This replaces mutation of underlying bytes in the iterated slice
with a shift counter, which is used when reading the head byte.
This is avoids having to copy the entire slice for every new iterator.
xor encoding is fast enough for our purposes and provides
very good compression.
We remove all other ones that partially don't support floats
for the sake of simplicity.
The adapter drags in a whole dependency chain of Prometheus internal
packages due to its storage's hideous structure.
With dependency flattening the adapter can actually be included into
Prometheus in a working way but it breaks package builds. Thus we
comment it out until we found a way to fix this contraption.
